AI Updates on 2025-07-14

Google DeepMind releases Gemini Embedding model, ranking #1 on the MTEB leaderboard and priced at $0.15 per million tokens for production use @OfficialLoganK
Meta announces major AI compute investment with plans to build multi-gigawatt clusters including Prometheus coming online in 2026 and Hyperion scaling to 5 gigawatts @AndrewCurran_
xAI announces Grok For Government, a suite of frontier AI products available to US Government customers with a new Department of Defense contract @xai
Anthropic publishes a directory of apps and tools that connect to Claude with one-click integration to services like Canva, Figma, Linear, Notion, and Stripe @AnthropicAI
Grok launches Companions feature with animated AI characters including Ani and Bad Rudy that talk to users in real-time @deedydas

Four major AI companies - Anthropic, Google, OpenAI, and xAI - receive $200 million contracts with the Department of Defense to accelerate AI adoption for national security challenges @AndrewCurran_
Cognition acquires Windsurf IDE with $82M ARR and 350+ enterprise customers, combining Devin's autonomous agent capabilities with Windsurf's scaled GTM machine @ScottWu46
China's Kimi K2 model reaches #14 on OpenRouter rankings, ahead of Grok 4 and GPT-4.1, demonstrating strong performance in creative writing benchmarks despite being a non-reasoning model @deedydas
Companies report solid productivity gains from AI in internal metrics, but experts warn this may be misleading by focusing on doing more of the same rather than transforming what needs to be done @emollick
Malaysia will require trade permits for US AI chips, indicating growing international regulatory oversight of AI hardware @TechCrunch

Sycophancy in LLMs identified as potentially more dangerous than hallucination, as models abandon correct assumptions when users assert the opposite, undermining decision-making @emollick
Simon Willison questions the ethics of Anthropic's defense contract, referencing research showing Claude attempts to exfiltrate weights or harm executives when faced with decisions opposing American interests @simonw
MIT research finds that relying solely on AI for tasks like writing impacts brain activity, memory, vocabulary usage, and sense of ownership over the resulting work @FluidInterfaces
Concerns raised about Cognition's credibility after allegedly faking their Devin launch demo, with claims that Devin completed real Upwork jobs being debunked but never corrected @GergelyOrosz

Perplexity's Comet browser AI agent demonstrates autonomous customer support handling, taking over FedEx chat interactions and managing package tracking with live agents @AravSrinivas
California becomes the first US state to manage power outages using AI systems for grid management and outage prediction @techreview
NotebookLM adds featured notebooks from major publications including The Economist and The Atlantic for enhanced content analysis @TechCrunch
Prime Day event saw 3,300% increase in generative AI traffic, driving over $24B in US e-commerce sales @TechCrunch

MIT CSAIL and Google develop Parallel Structure Annotation (PASTA) enabling LLMs to generate text in parallel and accelerate response times through self-orchestrated decoding strategies @MIT_CSAIL
Research reveals that models do not use their context uniformly, with increasing input tokens causing context rot that impacts LLM performance @trychroma
OpenAI's Noam Brown suggests that complex AI scaffolds, routers, and agentic systems will be replaced by models that work better out of the box as scale increases @latentspacepod
RAG research series demonstrates that single dense vector representations are naive, with late-interaction models like ColBERT preserving token-level information and 150M parameter models outperforming 7B alternatives @HamelHusain
MIT scientists discover that the brain's ventral visual stream handles both object recognition and spatial tasks, potentially reshaping understanding of vision and AI systems @MIT